Method and system for identifying and adding capacity of burst load in a computing power network

Abstract

The application belongs to the technical field of computing power network, and discloses a burst load discrimination and additional capacity expansion method in a computing power network, and proposes a burst recognition and adaptive expansion method: in the burst detection layer, the prediction residual of PACE is taken as the core signal, combined with an adaptive threshold, morphological constraints and pseudo-label semi-supervised self-training to improve cross-scene robustness; in the capacity disposal layer, the detection result is mapped to rapid additional expansion above the baseline, and through a step-by-step removal mechanism, the oscillation and cost waste in the falling back stage are inhibited. Through the normal state prediction control link of PACE and the abnormal state burst disposal link of BRACE, the stability-response-cost collaborative optimization under complex fluctuating load is realized, and the effectiveness of the proposed method in reducing invalid expansion and improving burst recognition reliability is verified in experiments.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of computing power network technology, and particularly relates to a method and system for identifying burst loads and scaling additional capacity in computing power networks. Background Technology

[0002] As computing networks evolve towards multi-domain collaboration across cloud, edge, and endpoint, distributed computing resources no longer exist solely as static resource pools. Instead, they require dynamic orchestration and elastic provisioning based on online business requests, regional traffic changes, and service quality constraints. Existing computing resource management typically relies on cloud-native container clusters, monitoring business load, resource utilization, or service metrics to trigger the addition, release, or migration of instance replicas, thus alleviating localized computing power shortages and resource idleness. Among these solutions, the closest technology is usually a resource management approach based on Kubernetes horizontal scaling or predictive scaling. Its basic logic involves collecting historical load or current monitoring metrics, determining the target number of instances based on fixed thresholds, predictive models, or resource estimation results, and then having the orchestration system execute scaling actions.

[0003] The above solutions can adapt to periodic business peaks or general load fluctuations to a certain extent. However, in real online business scenarios carried by computing networks, the load is often not a single stationary sequence, but rather a superposition of intraday periodic changes, slow trend drifts, and short-term high-frequency disturbances. For periodic and trend-based loads, the system needs to plan baseline capacity in advance to offset the time lag caused by instance startup, service readiness, and resource scheduling. For short-term spikes or abnormal disturbances, the system needs to identify risks and provide additional capacity within a short monitoring window. Existing single-path responsive scaling mainly relies on current metric out-of-bounds expansion, which is prone to starting expansion only after a sudden event, resulting in a lag in capacity compensation. While simple predictive scaling is forward-looking, predictive models are usually more suitable for characterizing repeatable regular changes. When external events, hotspot access, or local failures cause sudden load changes, prediction errors are directly transmitted to capacity decisions, resulting in undersupply or oversupply.

[0004] Furthermore, while existing burst-aware scaling solutions attempt to distinguish between periodic peaks and burst loads, their detection, discrimination, and handling stages are often relatively separate, and prediction residuals are not fully utilized as abnormal disturbance signals. Online detection often relies on fixed thresholds or short-window statistics, making it difficult to adapt to load baseline drift under low-latency conditions. The burst discrimination stage lacks a sample mechanism that can be automatically updated with the business status, easily misjudging normal peaks as abnormal bursts, or mistaking continuous abnormalities for trend changes. The capacity handling stage often treats expansion and contraction as symmetrical actions, lacking closed-loop control for rapid additional expansion and step-by-step release. Therefore, in the scenario of burst loads in computing power networks, a method that can coordinate normal capacity planning, residual disturbance detection, burst semantic discrimination, and additional capacity scaling is still needed. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method for identifying burst loads and scaling additional capacity in a computing network.

[0006] This invention is implemented as follows: A method for identifying burst loads and scaling additional capacity in a computing network includes:

[0007] Step 1: A dual-branch collaborative resource management framework based on load decomposition; through multi-angle perception and analysis of real business load, it was found that most online loads are more suitable to be represented as a composite structure of periodic changes + slow change trends + high-frequency fluctuation residuals.

[0008] Step 2: Divide the resource management process into a prediction-driven normal capacity planning link and an anomaly-driven emergency capacity expansion link, and connect the two links through prediction residuals.

[0009] Step 3: Using the prediction residuals of the normal link output as input for abnormal disturbances, the improved Welford online statistical algorithm is used in the detection layer to achieve low-latency dynamic standardization and preliminary screening of abnormal segments. In the discrimination layer, a semi-supervised self-training iterative framework based on pseudo-labels is introduced. In the capacity handling layer, a rapid additional capacity expansion and step-by-step release mechanism is further combined to achieve rapid capacity compensation and smooth recovery in sudden scenarios.

[0010] Furthermore, this involves multi-faceted perception and analysis of real business load:

[0011] A. Statistical distribution characteristics

[0012] The most basic quantity is the mean, which describes the average load level within the window:

[0013] (1)

[0014] Based on this, the maximum and minimum values , and range It is often used to reflect peak pressure and value range, making it easy to identify load objects that are generally high or have prominent peak values;

[0015] To measure the intensity of load fluctuations around the mean, the standard deviation is typically calculated. :

[0016] (2)

[0017] Considering that different objects may have different mean levels, only... Often incomparable, the coefficient of variation (CV) is frequently used in practice to characterize relative fluctuations. Here, we define CV as follows:

[0018] (3)

[0019] in It is a numerically stable term; generally speaking, The larger the value, the stronger the fluctuation at the unit load level, the worse the load stability, and the easier it is to induce frequent capacity adjustments at the control level.

[0020] When a load series exhibits peaks or tails, mean-variance statistics are easily affected by extreme points. Therefore, quantiles are often introduced to more robustly describe the distribution location: Let quantiles be... And the robust dispersion is characterized by the interquartile range:

[0021] (4)

[0022] Quantiles and IQR can eliminate the interference of a small number of extreme points to a certain extent. They are often used to compare the typical operating range and dispersion of different load objects, providing a basis for subsequent anomaly detection or threshold setting.

[0023] B. Trend and Slow-Change Characteristics

[0024] A typical approach to characterizing trends is to fit a window sequence using a univariate linear model:

[0025] (5)

[0026] in This represents the average rate of change of the load within the window over time. The intercept is... The remaining term not explained by the trend; slope This usually corresponds to a sustained increase in load. The corresponding load continues to decline; this quantity is often used in load analysis to distinguish between monotonic evolution and oscillating types without a clear direction, and to provide directional information for subsequent capacity planning;

[0027] Let the correlation coefficient of the regression fit be Then the trend strength index can be defined as

[0028] (6)

[0029] Building upon trend fitting, a crucial next step is detrending, which involves removing the portion of the original sequence that can be explained by the trend, resulting in the residual sequence.

[0030] (7)

[0031] The residual standard deviation can then be used to measure the magnitude of the remaining volatility after removing slow variations:

[0032] (8)

[0033] This quantity has direct meaning in load-aware analysis: if The load remains relatively large, indicating that fluctuations within the window are primarily driven by rapid ups and downs, spikes, or noise; trend modeling alone is insufficient to explain the load. A significant decrease indicates that the slow-changing trend accounts for a larger proportion of the load explanation, and the load has a stronger structural nature.

[0034] C. Features related to the rate of change;

[0035] D. Time-related features;

[0036] E. Frequency domain periodicity characteristics;

[0037] F. Complexity-related features.

[0038] Furthermore, the change rate type feature:

[0039] In load-aware analysis, differential classification features are typically introduced to measure the magnitude and direction of local changes in a sequence.

[0040] The most commonly used is the first-order difference sequence:

[0041] (9)

[0042] based on Three intuitive quantities can be obtained: its mean The standard deviation represents the overall direction of change within the characterizing window. The mean absolute difference characterizes the stability of the rate of change. This depicts the typical amplitude of the jump;

[0043] Define the number of ascents and the number of descents:

[0044] (10)

[0045] And can be further used in proportion Describe whether rising or falling is dominant within the window.

[0046] Furthermore, the time-related features

[0047] Lag Autocorrelation can be defined as:

[0048] (11).

[0049] Furthermore, the frequency domain periodicity class feature:

[0050] After removing the mean from the window sequence, its Discrete Fourier Transform (DFT) can be constructed.

[0051] (12)

[0052] With amplitude spectrum Describe the intensity of each frequency component; locate the main peak frequency within the positive frequency range:

[0053] (13)

[0054] (14).

[0055] Furthermore, the complexity class features:

[0056] Sample entropy is a commonly used metric in time series complexity analysis. It reflects the reproducibility of local patterns by comparing the pattern-matching probabilities of sequences under different embedding dimensions; let the embedding dimension be... The construction length is Template vector:

[0057] (15)

[0058] The similarity between different templates is measured using some distance metric (commonly Chebyshev distance), and the statistical distance is less than the tolerance. The number of matching pairs yields the dimension. Matching ratio below And then expand to dimensions get Sample entropy is defined as:

[0059] (16).

[0060] Another objective of this invention is to provide a system for identifying and scaling additional capacity for burst loads in a computing network, comprising:

[0061] The perception and analysis module is used for a dual-branch collaborative resource management framework based on load decomposition. By perceiving and analyzing real business loads from multiple perspectives, it is found that most online loads are more suitable to be represented as a composite structure of periodic changes + slow change trends + high-frequency fluctuation residuals.

[0062] The partitioning module is used to divide the resource management process into a prediction-driven normal capacity planning link and an anomaly-driven emergency capacity expansion link, and connect the two links through prediction residuals.

[0063] The filtering module uses the prediction residuals of the normal link output as input for abnormal disturbances. In the detection layer, it uses an improved Welford online statistical algorithm to achieve low-latency dynamic standardization and preliminary screening of abnormal segments. In the discrimination layer, it introduces a semi-supervised self-training iterative framework based on pseudo-labels. In the capacity handling layer, it further combines a rapid additional expansion and step-by-step release mechanism to achieve rapid capacity compensation and smooth recovery in sudden scenarios.

[0064] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the method for identifying burst loads and additional capacity scaling in the computing network.

[0065] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of a method for identifying burst loads and additional capacity scaling in the computing network.

[0066] Another objective of this invention is to provide an information data processing terminal, which is used to implement a system for identifying burst loads and adding capacity scaling in the computing network.

[0067] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0068] This invention first conducts perception analysis and pattern summarization on real-world business loads. The results show that strictly speaking, stable cyclical patterns that repeat at fixed times are not common; most online loads are better characterized as a composite structure of a cyclical skeleton, a slow drift term, and a high-frequency residual signal. The cyclical skeleton and drift term reflect predictable load changes, suitable for establishing baseline capacity and providing decision lead time; the high-frequency residual signal characterizes abnormal disturbances that deviate from the normal evolutionary trajectory, serving as an important basis for identifying sudden risks.

[0069] Based on the above analysis, this invention proposes a dual-branch collaborative resource management framework based on load decomposition. This framework divides the resource regulation process into two interconnected links: one is a prediction-driven link oriented towards normal load changes, used for baseline capacity planning and forward-looking elastic control based on periodic and trend information; the other is an anomaly-driven link oriented towards abnormal disturbances, used for identifying sudden events based on residual out-of-bounds signals and performing rapid additional capacity expansion. By using the prediction residual as the connection signal between the two links, the framework achieves a coordinated unification of prediction-driven forward-looking regulation and anomaly-driven fallback handling. This invention further elaborates on the method design along the two links of the aforementioned dual-branch framework.

[0070] To address the issues of capacity mismatch and scaling oscillations that easily occur under normal load variations, this invention proposes a predictive-driven elastic scaling method, PACE (Proactive Autoscaling with Cyclic Load Estimation), on the predictive-driven link of the aforementioned dual-branch framework. This method constructs a rolling short-term prediction model based on load decomposition results, uses a cyclic skeleton and trend information to form a baseline capacity, and designs an elastic control strategy of advance expansion and phased recovery accordingly. This allows the system to proactively warm up and expand capacity during load increases by leveraging predictive lead time, and smoothly recover redundant capacity during load decreases, thereby reducing the impact of container startup latency on tail performance and mitigating system oscillations caused by ineffective expansion and contraction. Experimental results show that PACE can reduce the SLO default rate by 17.7-23.0 percentage points compared to HPA while ensuring service stability.

[0071] To address the issue of rapidly expanding capacity gaps and escalating SLO default risks under sudden traffic surges, this invention proposes a Burst Recognition and Adaptive Capacity Expansion (BRACE) method on the anomaly-driven link of the aforementioned dual-branch framework. This method uses the aforementioned prediction residuals as input and residual out-of-bounds signals as entry points. It combines adaptive thresholds and morphological constraints to perform initial screening of anomalous segments and further introduces a pseudo-label-based semi-supervised self-training mechanism to improve the robustness of burst detection under label-scarce conditions. At the capacity handling layer, BRACE maps the detection results to rapid additional capacity expansion above the baseline, and combines minimum duration, capacity upper limit, and step-by-step rollback strategies to achieve smooth release. This rapidly suppresses service risks during sudden disturbances while controlling resource overhead. Experimental results show that BRACE can suppress instantaneous default traffic at the peak of a burst from 500–800 rps to near zero, thus significantly reducing the SLO default risk in burst scenarios.

[0072] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0073] Given the current large-scale construction of computing infrastructure and the growing demand for cross-regional collaboration, the industry is emphasizing a computing power interconnection system oriented towards cloud-edge-device collaboration to alleviate the mismatch between computing power supply and demand and business carrying capacity. In the context of the coexistence of multi-source heterogeneous computing power and the continuous expansion of business distribution, it is necessary to improve the perceptibility, connectivity, and schedulability of computing resources to support efficient allocation and application deployment. The goals of schedulability, rapid provisioning, and unified governance typically rely on cloud-native technology systems. While cloud-native container clusters possess engineering advantages such as on-demand provisioning and elastic recovery, their actual effectiveness largely depends on whether the system can provide timely and stable capacity within critical time windows of load changes. Therefore, this invention provides a feasible solution for further cost control. The application documents of this invention are adapted to the layout requirements of the current standardization stage, match the patent strategies of the early / mid / late stages in the guidelines, and reserve space for the application of priority, novelty grace period, and deferred examination systems, complying with the patent application strategy requirements of the entire standardization process in the guidelines.

[0074] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0075] To address the challenge of simultaneously managing both forward-looking planning for normal loads and rapid backup for sudden loads in cloud-native container clusters under complex, time-varying loads, this invention proposes a dual-branch collaborative resource management framework based on load decomposition, focusing on the synergistic relationship between business load awareness, elastic control, and resource optimization within the context of computing power networks. To address the issue of rapidly expanding capacity gaps and significantly increased SLO default risk under sudden load disturbances, a sudden load identification and adaptive scaling method, BRACE, is proposed. This method uses the prediction residuals from normal link outputs as input for abnormal disturbances. At the detection layer, an improved Welford online statistical algorithm is used to achieve low-latency dynamic standardization and initial screening of abnormal segments. At the discrimination layer, a semi-supervised self-training iterative framework based on pseudo-labels is introduced to improve the robustness and generalization ability of sudden load identification under label scarcity conditions. At the capacity handling layer, a rapid additional scaling and step-by-step release mechanism is further combined to achieve rapid capacity compensation and smooth recovery in sudden load scenarios. Thus, this invention constructs an integrated method chain for rapid detection, accurate discrimination, and stable handling of sudden loads. Attached Figure Description

[0076] Figure 1 This is a flowchart of the method for identifying sudden loads and scaling additional capacity in a computing network provided in an embodiment of the present invention.

[0077] Figure 2This is a block diagram of the burst load discrimination and additional capacity scaling system in the computing power network provided in the embodiments of the present invention.

[0078] Figure 3 This is a diagram showing the current state of expansion and contraction provided in an embodiment of the present invention.

[0079] Figure 4 This is a MAPE-K architecture diagram provided in an embodiment of the present invention.

[0080] Figure 5 This is a diagram of a dual-branch resource management framework provided in an embodiment of the present invention.

[0081] Figure 6 This is a flowchart of the streaming residual screening process based on Welford online statistics provided in an embodiment of the present invention.

[0082] Figure 7 This is a diagram of a semi-supervised self-training iterative framework based on confidence awareness provided in an embodiment of the present invention.

[0083] Figure 8 This is a graph showing the performance variation of semi-supervised iteration provided in an embodiment of the present invention.

[0084] Figure 9 This is a test set confusion matrix diagram provided in an embodiment of the present invention.

[0085] Figure 10 This is the feature distribution and approximate discriminant diagram under PCA dimensionality reduction provided in this embodiment of the invention.

[0086] Figure 11 This is a comparison chart of capacity supply and load demand under different strategies provided in the embodiments of the present invention.

[0087] Figure 12 This is a comparison chart of instantaneous defaults provided in an embodiment of the present invention.

[0088] Figure 13 This is a graph showing the cost rate changing over time (0.0014 yuan / minute / instance) provided in an embodiment of the present invention.

[0089] Figure 14 This is a graph showing the cumulative expenditure difference relative to BRACE, provided in an embodiment of the present invention.

[0090] Figure 15 This is a graph showing the change in the number of active instances provided in an embodiment of the present invention.

[0091] Figure 16 This is a comparison chart of resource utilization rates for various strategies provided in the embodiments of the present invention.

[0092] Figure 17 This is a graph showing the change in idle capacity provided in an embodiment of the present invention. Detailed Implementation

[0093] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The embodiments of this invention provide a method and system for identifying burst loads and scaling additional capacity in a computing network. Those skilled in the art, following the description, can implement this technical solution, solve its technical problems, and achieve the expected technical effects.

[0094] like Figure 1 As shown in the figure, the method for identifying burst loads and scaling additional capacity in a computing network provided by this embodiment of the invention includes the following steps:

[0095] S101 is a dual-branch collaborative resource management framework based on load decomposition. Through multi-angle perception and analysis of real business load, it is found that most online loads are more suitable to be represented as a composite structure of periodic changes + slow change trends + high-frequency fluctuation residuals.

[0096] S102 divides the resource management process into a prediction-driven normal capacity planning link and an anomaly-driven emergency capacity expansion link, and connects the two links through prediction residuals.

[0097] S103 uses the prediction residuals of the normal link output as input for abnormal disturbances. In the detection layer, it uses an improved Welford online statistical algorithm to achieve low-latency dynamic standardization and preliminary screening of abnormal segments. In the discrimination layer, it introduces a semi-supervised self-training iterative framework based on pseudo-labels. In the capacity handling layer, it further combines a rapid additional expansion and step-by-step release mechanism to achieve rapid capacity compensation and smooth recovery in sudden scenarios.

[0098] like Figure 2 As shown in the figure, an embodiment of the present invention provides a system for identifying burst loads and scaling additional capacity in a computing network, comprising:

[0099] The perception and analysis module is used for a dual-branch collaborative resource management framework based on load decomposition. By perceiving and analyzing real business loads from multiple perspectives, it is found that most online loads are more suitable to be represented as a composite structure of periodic changes + slow change trends + high-frequency fluctuation residuals.

[0100] The partitioning module is used to divide the resource management process into a prediction-driven normal capacity planning link and an anomaly-driven emergency capacity expansion link, and connect the two links through prediction residuals.

[0101] The filtering module uses the prediction residuals of the normal link output as input for abnormal disturbances. In the detection layer, it uses an improved Welford online statistical algorithm to achieve low-latency dynamic standardization and preliminary screening of abnormal segments. In the discrimination layer, it introduces a semi-supervised self-training iterative framework based on pseudo-labels. In the capacity handling layer, it further combines a rapid additional expansion and step-by-step release mechanism to achieve rapid capacity compensation and smooth recovery in sudden scenarios.

[0102] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the method for identifying burst loads and additional capacity scaling in the computing network.

[0103] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of a method for identifying burst loads and additional capacity scaling in the computing network.

[0104] Another objective of this invention is to provide an information data processing terminal, which is used to implement a system for identifying burst loads and adding capacity scaling in the computing network.

[0105] Specific implementation of the present invention:

[0106] Figure 3 Current status of expansion

[0107] 1. To address the challenge of simultaneously managing both proactive planning for normal loads and rapid backup for sudden loads under complex, time-varying loads in cloud-native container clusters, this invention proposes a dual-branch collaborative resource management framework based on load decomposition, within the context of computing power networks, focusing on the synergistic relationship between business load perception, elastic control, and resource optimization. It should be noted that the resources discussed in this invention primarily refer to the elastic execution resources in cloud-native container clusters for online service operation, specifically manifested as horizontal scaling capabilities centered on the number of container instances, and associated with the basic resource configurations of individual instances, such as CPU and memory. This invention no longer treats complex loads as a single, uniform object, but rather divides them into normal loads dominated by periodic and slow-changing trends, and abnormal loads characterized by short-term, drastic fluctuations, designing corresponding resource regulation mechanisms for each, thereby improving service quality assurance capabilities and resource utilization efficiency.

[0108] First, addressing the challenge of existing resource management methods simultaneously handling predictable normal changes and sudden, abnormal disturbances, this invention proposes a dual-branch collaborative resource management framework based on load decomposition. Through multi-faceted perception and analysis of real business loads, it was found that most online loads are better represented as a composite structure of periodic changes, slow trends, and high-frequency fluctuation residuals. Based on this, the invention divides the resource management process into a prediction-driven normal capacity planning link and an anomaly-driven emergency capacity expansion link, connecting the two links through prediction residuals. This provides a unified methodological foundation for subsequent normal load regulation and emergency load handling.

[0109] Secondly, addressing the issue of rapidly expanding capacity gaps and significantly increased SLO default risk under sudden load disturbances, this invention proposes the BRACE method for sudden load identification and adaptive expansion. This method uses the prediction residuals from normal link outputs as input for abnormal disturbances. At the detection layer, an improved Welford online statistical algorithm is used to achieve low-latency dynamic standardization and preliminary screening of abnormal segments. At the discrimination layer, a pseudo-label-based semi-supervised self-training iterative framework is introduced to improve the robustness and generalization ability of sudden load identification under label scarcity conditions. At the capacity handling layer, a rapid additional expansion and step-by-step release mechanism are further combined to achieve rapid capacity compensation and smooth recovery under sudden load scenarios. Thus, this invention constructs an integrated method chain for rapid detection, accurate discrimination, and stable handling of sudden loads.

[0110] 2. Theoretical Foundations of the General Resource Management Framework

[0111] Existing resource management frameworks typically follow a classic autonomous control closed-loop architecture, with MAPE (Monitor–Analyze–Plan–Execute) as the main thread, supplemented by a knowledge base to form an autonomous manager (AutonomicManager) of MAPE-K. This framework aims to achieve automated management and closed-loop optimization of computing resources by continuously sensing system status, generating resource decisions online, and executing adjustment actions.

[0112] like Figure 4As shown, the autonomous manager first continuously collects business and system status indicators through the monitoring module, forming a unified observation of managed resources (such as cloud servers, networks, and storage). Then, the analysis module performs online diagnostics and status assessments on the observed data to determine whether the current resource supply can meet load demands or whether there are any abnormal risks. Based on this, the planning module calculates resource adjustment schemes according to policy rules or models, generating the next executable control action. Finally, the execution module translates the plan into specific resource adjustment instructions and completes scaling up / down, migration, or parameter updates through the orchestration system interface, thus returning the system status to the monitoring end to enter the next closed loop. A knowledge base spanning all modules is used to accumulate historical observations, policies, and model parameters, providing a basis for analysis and planning decisions and supporting continuous iteration and updates of policies.

[0113] In cloud-native container clusters, online service operation generates continuously changing business loads at the system level, and the effectiveness of resource management largely depends on the degree of understanding of these load changes. Real business loads are neither completely random fluctuations nor strictly repeating ideal periodic signals, but rather contain both perceptible structural patterns and non-negligible random disturbances over a long period of evolution. Ignoring this complexity and directly adjusting resources based on instantaneous states can easily lead to capacity supply lags, scaling oscillations, and service quality degradation. Therefore, it is necessary to conduct a systematic perception analysis of the business load before designing the resource management framework, thereby providing a basis for subsequent load decomposition, capacity planning, and elastic control.

[0114] It should be noted that the term "business" in this invention primarily refers to online services and their business scenarios running in cloud-native container clusters, while "load" refers to the observable resource demands and operational pressure exerted on the system by the business during operation, typically manifested as time series such as request volume, memory usage, and queue length. The two are not equivalent. Businesses themselves do not necessarily exhibit strictly repetitive cyclical patterns, and different businesses differ significantly in access methods, user groups, and operating environments. However, from a system observation perspective, the load sequence formed by business operations is often influenced by factors such as user access frequency, platform operating mechanisms, and external events, thus exhibiting certain patterns in a statistical sense, further manifesting as a cyclical structure, a slow changing trend, and the superposition of random disturbances.

[0115] Based on this understanding, this invention defines load perception as: the entire process of acquiring, characterizing, analyzing, and identifying the statistical characteristics, temporal patterns, evolutionary laws, and abnormal states of business load, starting from system monitoring data, and providing a basis for subsequent resource regulation decisions. Its connotation includes not only feature extraction, pattern analysis, and structural decomposition of historical load, but also forward-looking modeling for normal load changes and identification and discrimination of abnormal disturbances. In other words, load prediction is a further application of load perception in normal scenarios, while anomaly identification is a deeper extension of load perception in abnormal scenarios; both serve capacity planning and elastic control within the resource management framework.

[0116] Therefore, before proceeding with the normal prediction-driven scaling method and the burst identification-driven expansion method, it is necessary to first conduct a systematic perception analysis of the business load based on historical monitoring data, clarify its main statistical characteristics, temporal patterns, and structural composition, and further construct a matching resource management framework accordingly. Based on this, this chapter first establishes a load characteristic characterization method and indicator system to conduct multi-dimensional analysis of real business loads; then, combining empirical results, it performs structural decomposition of complex loads, dividing them into normal variation components and abnormal disturbance components; finally, based on this, it designs a dual-branch resource management framework to provide a unified modeling foundation and architectural support for the subsequent PACE and BRACE methods.

[0117] 3.1 Load Awareness Methods and Metric Definitions

[0118] To identify the main temporal patterns of business load from historical monitoring data and provide a basis for subsequent load decomposition and resource regulation, it is first necessary to systematically characterize and analyze the load sequences. This invention uses windowed load sequences as the basic analysis object and employs a multi-view feature extraction + rule-based pattern discrimination method for load perception. The constructed feature system covers seven types of information: statistics, trends, changes, autocorrelation, frequency domain, complexity, and abrupt changes.

[0119] A. Statistical distribution characteristics

[0120] When performing perception analysis on historical loads, the process typically begins with the overall distribution pattern and then analyzes the window load sequence. We extract a set of classic statistics to characterize the average level, dispersion, and tail shape of the load. These indicators do not rely on specific model assumptions, are simple to calculate, highly interpretable, and suitable for rapid scanning and cross-sectional comparisons on large-scale samples.

[0121] The most basic quantity is the mean, which describes the average load level within the window:

[0122] (1)

[0123] Based on this, the maximum and minimum values , and range It is often used to reflect peak pressure and value range, making it easy to identify load objects that are generally high or have prominent peak values.

[0124] To measure the intensity of load fluctuations around the mean, the standard deviation is typically calculated. :

[0125] (2)

[0126] Considering that different objects may have different mean levels, only... Often incomparable, the coefficient of variation (CV) is frequently used in practice to characterize relative fluctuations. Here, we define CV as follows:

[0127] (3)

[0128] in This is a numerically stable term. Generally speaking, The larger the value, the stronger the fluctuation at the unit load level, the worse the load stability, and the easier it is to induce frequent capacity adjustments at the control level.

[0129] When a load series exhibits peaks or tails, mean-variance statistics are easily affected by extreme points. Therefore, quantiles are often introduced to more robustly describe the distribution location: Let quantiles be... And the robust dispersion is characterized by the interquartile range:

[0130] (4)

[0131] Quantiles and IQR can eliminate the interference of a small number of extreme points to a certain extent. They are often used to compare the typical operating range and dispersion of different load objects, providing a basis for subsequent anomaly detection or threshold setting.

[0132] Furthermore, skewness and kurtosis can be used to supplement information about distribution patterns: skewness characterizes the asymmetry of the distribution, while kurtosis reflects the peak size and tail variation characteristics. In load analysis, skewness and kurtosis are often used to identify typical patterns of low load most of the time with peaks occurring only occasionally, statistically indicating the need to pay attention to the impact of tail risks on performance and capacity decisions.

[0133] B. Trend and Slow-Change Characteristics

[0134] In addition to the overall distribution pattern, real business load often includes slow changes over time, such as continuous rises / falls within a window, and slow increases / decreases in the baseline level. To quantify these slow-changing structures statistically, a common practice is to introduce a trend term into the window sequence and separate it from the fast-fluctuation portion, thereby determining whether the changes are mainly driven by interpretable slow variations and how strong the remaining fluctuations are after removing the slow variations.

[0135] A typical approach to characterizing trends is to fit a window sequence using a univariate linear model:

[0136] (5)

[0137] in This represents the average rate of change of the load within the window over time. The intercept is... This represents the remaining term not explained by the trend. Slope This usually corresponds to a sustained increase in load. The corresponding load continues to decline; this quantity is often used in load analysis to distinguish between monotonic evolution and oscillating types without a clear direction, and to provide directional information for subsequent capacity planning.

[0138] The magnitude of the slope alone is insufficient to determine whether a trend is stable; therefore, the correlation coefficient of the regression fit is often used to measure the strength of the trend. Let the correlation coefficient of the regression fit be... Then the trend strength index can be defined as

[0139] (6)

[0140] The larger the value, the stronger the linear correlation between the sequence and the linear time axis, indicating a more consistent trend and a closer approximation to an interpretable slow-changing structure; conversely, a smaller value indicates a weaker correlation. A small slope, even if not zero, may indicate that the sequence fluctuates wildly within the window, making the trend difficult to describe stably. In engineering applications, trend strength can be used to determine whether current load changes are primarily slow and predictable, or more influenced by short-term fluctuations.

[0141] Building upon trend fitting, a crucial next step is detrending, which involves removing the portion of the original sequence that can be explained by the trend, resulting in the residual sequence.

[0142] (7)

[0143] The residual standard deviation can then be used to measure the magnitude of the remaining volatility after removing slow variations:

[0144] (8)

[0145] This quantity has direct meaning in load-aware analysis: if The load remains relatively large, indicating that fluctuations within the window are primarily driven by rapid ups and downs, spikes, or noise; trend modeling alone is insufficient to explain the load. A significant decrease indicates that the slow-changing trend accounts for a larger proportion of the load explanation, and the load has a stronger structural nature. Through the link of trend fitting-detrending-residual fluctuation, slow and fast changes can be initially separated at the statistical level, laying the foundation for subsequent characterization of periodicity, mutation, and complexity.

[0146] C. Characteristics related to rate of change

[0147] Distribution statistics alone are insufficient to reflect the rate of load change over time, a property that scaling control is highly sensitive to: given the same mean and variance, a faster rate of load change often triggers frequent scaling or control jitter. Therefore, load-aware analysis typically incorporates differential classification features to measure the magnitude and direction of local changes in the sequence.

[0148] The most commonly used is the first-order difference sequence:

[0149] (9)

[0150] based on Three intuitive quantities can be obtained: its mean The standard deviation represents the overall direction of change within the characterizing window. The mean absolute difference characterizes the stability of the rate of change. This depicts the typical amplitude of abrupt changes. Compared to simply looking at... Fluctuations in these differential indicators more closely reflect the action pressure of the control layer: when or When the value is large, it indicates that the load fluctuates more drastically between adjacent time periods. If the control strategy lacks stability constraints such as hysteresis / speed limiting, it is usually more likely to make frequent adjustments.

[0151] Furthermore, the sign of the difference can provide supplementary information about the bias of change. Define the number of ascents and descents:

[0152] (10)

[0153] And can be further used in proportion This describes whether the upward or downward trend dominates within the window. This type of quantity is often used to distinguish between continuously climbing loads and high-frequency oscillating loads, and provides supplementary evidence for separating subsequent trends / drifts from anomalous disturbances.

[0154] D. Time-related features

[0155] Real-world business workloads often exhibit a time-dependent structure, such as short-term inertia, repetitive patterns, or periodic skeletons. To quantify this temporal repeatability, load-aware analytics typically calculates the autocorrelation function (ACF) to characterize the correlation of sequences at different lags. Autocorrelation can be defined as:

[0156] (11)

[0157] when Maintaining a high value within a small lag range indicates that the sequence has stronger short-term inertia and continuity, and the load is smoother between adjacent time points; when Rapid decay approaching zero indicates that the sequence is closer to random fluctuations or is dominated by high-frequency disturbances. Significant correlation peaks on certain lags may suggest the presence of recurring structures, such as periodic or quasi-periodic patterns.

[0158] From an engineering perspective, autocorrelation features are often used to answer two key questions: first, whether the load has stable short-term dependencies, thus providing a usable structure for short-term prediction or control; and second, whether the load exhibits recurring patterns, making methods based on historical templates reasonable. Complementing differential features, ACF emphasizes structure and repeatability, while differential features emphasize the rate of change and control sensitivity.

[0159] E. Frequency domain periodicity characteristics

[0160] Periodicity is one of the most common and easily exploitable structures in workloads. However, in real-world data, periodic signals are often masked by trend drift and high-frequency disturbances, making intuitive observation in the time domain unreliable. Frequency domain methods, by decomposing the sequence into different frequency components, can more directly answer questions such as whether a dominant periodic scale exists and whether periodic energy is concentrated. Therefore, they are often used for periodicity testing and coarse-grained periodic intensity measurement in the load sensing phase.

[0161] After removing the mean from the window sequence, its Discrete Fourier Transform (DFT) can be constructed.

[0162] (12)

[0163] With amplitude spectrum Describe the intensity of each frequency component. Locate the dominant peak frequency within the positive frequency range:

[0164] (13)

[0165] This provides clues to the periodic scale corresponding to the dominant frequency: the lower the dominant frequency, the longer the corresponding period; the higher the dominant frequency, the shorter the corresponding period. To further measure the prominence of the periodic component, indicators such as the proportion of dominant frequency energy are often introduced, for example, using the ratio of the peak amplitude to the total amplitude to approximate the energy concentration.

[0166] (14)

[0167] when A higher frequency indicates that the sequence energy is more concentrated at a certain dominant frequency, and the periodic structure is more pronounced; when... At lower values, it indicates that the spectral energy is more dispersed, possibly dominated by multiple period superpositions, drift disturbances, or high-frequency noise. The advantage of this type of frequency domain feature is that it provides a direct quantitative result of whether the period is prominent, and it can be mutually verified with the autocorrelation feature: a strong ACF but a lack of prominent dominant frequency often corresponds to short-term inertia or quasi-periodicity; a prominent dominant frequency is more likely to have a reusable periodic skeleton.

[0168] F. Complexity-related features

[0169] Besides exhibiting periodicity and trends, load sequences may also display varying degrees of regularity and randomness. This dimension is equally crucial for scaling and prediction: sequences with strong regularity are more predictable and better suited for constructing forward-looking strategies; sequences with strong randomness rely more on robust control and anomaly fallback. The goal of complexity metrics is to statistically measure the degree of repetition of sequence patterns, thereby reflecting their predictability.

[0170] Sample entropy is a commonly used metric in time series complexity analysis. It reflects the reproducibility of local patterns by comparing the pattern-matching probabilities of sequences under different embedding dimensions. Let the embedding dimension be... The construction length is Template vector:

[0171] (15)

[0172] The similarity between different templates is measured using some distance metric (commonly Chebyshev distance), and the statistical distance is less than the tolerance. The number of matching pairs yields the dimension. Matching ratio below And then expand to dimensions get Sample entropy is defined as:

[0173] (16)

[0174] Its intuitive meaning is: if the length is Once the local pattern is matched, it extends to The higher the probability of maintaining a match, the more regular the sequence and the lower the sample entropy; conversely, if the local pattern is difficult to sustain stably, the sequence is more complex and the sample entropy is greater. In load perception analysis, sample entropy is often used to distinguish between predictable loads with stable and repetitive structures and highly uncertain loads driven by disturbances. Together with periodicity (dominant frequency energy) and correlation (ACF), it forms the basis for a multidimensional characterization of the interpretable structure of the load.

[0175] 3.2 Dual-branch resource management framework based on load decomposition

[0176] 3.2.1 Load Decomposition Modeling

[0177] Considering the discrete time axis Recording cloud-native business at all times The observation load is To characterize the composite features of a real load having a periodic structure but not strict repetition, accompanied by slow shifts and high-frequency disturbances, this invention represents the load sequence as a superposition of three terms:

[0178] (17)

[0179] in, This indicates the cyclical contribution dominated by repetitive structures such as intraday rhythms; This indicates a slower drift at a lower frequency relative to the period, used to describe gradual changes in the baseline level; The residual term aggregates high-frequency fluctuations and anomalous disturbances that are difficult to explain by periodic structure and slow drift.

[0180] To address the trans-day waveform variations and significant phase shifts observed in Section 3.2, a rigid absolute period assumption cannot be applied. Therefore, the periodic structure in this section is represented as a quasi-periodic form using a periodic template and slow phase / amplitude modulation. Let the period length be... Define phase variable And introduce a periodic template mapping (Only depicting intraday shapes), satisfying

[0181] (18)

[0182] Considering that the peak time and peak height of real business will change slowly over time, this invention uses a slowly changing amplitude coefficient. With phase shift Modulate the template to define the periodic term as follows:

[0183] (19)

[0184] in Used to characterize the slow fluctuations in periodic amplitude. This is used to characterize the slow left-right drift of periodic peaks and troughs on the time axis. To ensure that the above modulation only carries gradual changes and does not absorb short-term spikes or high-frequency oscillations, this invention... and Apply slow variation constraints:

[0185] (20)

[0186] in Used to balance the dimensional differences between amplitude variation and phase drift within the same constraint. This is the upper bound of the total slowly variable allowed for the quasi-periodic term. Besides the periodic component, the drift term... To characterize low-frequency non-stationarity beyond periodic contributions, it can include baseline bias and long-term trends. Similarly, to avoid misclassifying short-term spikes as drift, this invention constrains its gradual variability with bounded total variation:

[0187] (twenty one)

[0188] From equation (3-17), the residual term can be defined as follows:

[0189] (twenty two)

[0190] At this time Random fluctuations are used to represent the complexity of sequences.

[0191] 3.2.2 Overall Architecture of Elastic Resource Management Framework

[0192] Based on load awareness analysis, stable cycles that repeat exactly at the same time every day are not common in real business loads. A more reasonable characterization is the superposition of normal evolution dominated by the cycle skeleton and slow drift terms, and abnormal disturbances represented by high-frequency residuals. Based on the above understanding and the load decomposition method in Section 3.3.1, this invention designs the resource management framework as two complementary functional branches: one for proactive resource adjustment of predictable cycle and slow drift components, and the other for rapid perception and dynamic response to sudden disturbances and abnormal fluctuations, thus forming a dual-branch collaborative management mechanism for complex load scenarios. The overall architecture of the entire elastic resource management framework is as follows: Figure 5 As shown, from left to right, it can be divided into an input layer, a processing layer, and a decision fusion and execution layer.

[0193] First, the input layer is separated from the load. This layer directly interfaces with the monitoring system of the cloud-native cluster and is responsible for acquiring real-time raw load observation data. Then, through the load decomposition mechanism defined in Section 3.3.1, the raw time series is decomposed into structured normal components (periodic terms and drift terms) and unstructured residual input.

[0194] Secondly, there is the core processing layer, which includes a dual-link collaborative processing mechanism. The first is the PACE normal link. This link receives normal components and aims to predict the periodic term and drift term separately for the predictable changes corresponding to the periodic skeleton and drift, and finally synthesizes the prediction baseline and outputs the baseline capacity. Its core value lies in providing capacity expansion lead time, supporting forward-looking resource supply and steady-state maintenance. The second is the BRACE abnormal link. Faced with sudden flow or abnormal disturbances caused by residual out-of-bounds errors, single predictions often fail and are lagging. Therefore, this link uses residual input for streaming anomaly screening (based on the Welford algorithm to update statistics in real time) and combines semi-supervised classification for robust burst discrimination. Once an anomaly is confirmed, an aggressive capacity decision is immediately generated and additional capacity is output.

[0195] Finally, there is the decision fusion and execution layer. In this layer, the baseline capacity output from the PACE normal link is used to form the basic resource configuration of the system under normal operating conditions and enters the capacity coordination and fusion stage; the additional capacity output from the BRACE link is superimposed or released through additional capacity attenuation control of the baseline capacity. Subsequently, the baseline capacity configuration results and the additional capacity control results work together to the orchestration execution layer, driving the actual scaling up and down and resource orchestration operations, thereby forming a closed-loop resource management process for complex load scenarios.

[0196] In summary, this framework ensures both resource utilization and control stability under normal conditions, while also taking into account service reliability under extreme emergencies.

[0197] 4.1 Burst Load Characteristics Analysis and Problem Definition

[0198] This invention, based on a load decomposition-dual-track collaborative framework, divides real-world business loads into predictable normal components and unpredictable abnormal disturbances. The baseline capacity output from the PACE normal link is used for forward-looking planning. Unlike normal evolution, real-world production environments still experience widespread sudden traffic surges caused by promotional activities, hot events, or internal retry storms. These surges typically manifest as rapid increases and strong non-stationary fluctuations within minutes, making the baseline capacity, which relies solely on prediction, prone to significant gaps in the early stages of a surge, triggering queuing backlogs, soaring overall latency, and even SLO (Solution Logical Level) defaults. Against this backdrop, this chapter, building upon the normal baseline provided by PACE, further provides a formal definition of a surge and models resource supply decisions in surge scenarios as a joint optimization problem of performance risk and economic cost.

[0199] Based on the load composite structure model established above, the normal baseline load has been taken over by the PACE method. Therefore, burst identification no longer focuses on the absolute value of the total load, but rather on the residual portion that the normal model fails to explain. (Definition of time point) The predicted residual is

[0200] (twenty three)

[0201] in For actual observation load, For PACE at any time The output baseline prediction. Ideally, if the load follows normal evolution, the residual... It should be a controlled fluctuation with an approximate zero mean; however, when a sudden event occurs, They often exhibit significant and persistent positive deviations. Therefore, this chapter defines sudden events as: predictions of residual sequences. This approach identifies anomalous segments that simultaneously satisfy statistical significance in both amplitude and duration dimensions. It isolates burst detection from the original load, thus avoiding systematic interference from periodic fluctuations and long-term trends on the detection threshold, allowing the detection mechanism to focus on extreme perturbations that cannot be explained by the normal model.

[0202] After defining the burst signal, the core question the resource supply side needs to answer is: how much computing resources should be added when a burst occurs to achieve the optimal trade-off between the risk of service quality breach and the cost of resource over-provisioning? Continuing with the symbol definition from Chapter 3, we will record the time. The baseline replica count planned by PACE normal link is Its corresponding system baseline carrying capacity is ,in This represents the upper limit of processing capacity for a single container instance while satisfying the SLO constraint. Based on this, the additional capacity that the BRACE link needs to decide is defined as follows: Then the system at time... Total supply capacity The sum of the two:

[0203] (twenty four)

[0204] Further time setting The actual system load requirement is .because During emergencies, abnormal components are often difficult for normal models to capture, and their values ​​are frequently significantly higher than the demand levels implied by baseline predictions. Therefore, it is necessary to... To fill the supply-side gap in a short period of time.

[0205] To quantify the trade-off between gap risk and resource costs, this chapter constructs the following two types of cost functions. First, when the total supply capacity... Less than the actual load requirement At times, capacity shortages will lead to queuing or dropping, causing a spike in response time or an increase in error rate. Considering that service quality degradation under sudden high concurrency scenarios usually exhibits a non-linear avalanche effect, this invention uses a squared form to penalize insufficient capacity:

[0206] (25)

[0207] in This is a default risk penalty coefficient used to characterize the business's aversion to capacity shortages. Secondly, to prevent the system from requesting resources without limit in pursuit of stability, a resource holding cost is introduced: Let the operating cost of a unit container instance per unit of time be... The cost of adding a copy is

[0208] (26)

[0209] Combining the above two points, the decision objective for sudden adaptive capacity expansion is within the sudden sustained window. Within, find the optimal sequence of additional capacity. To minimize the total cumulative cost of the system:

[0210] (27)

[0211] The essence of this optimization problem lies in balancing the relative strengths of the default risk term and the resource cost term: when When the capacity is relatively large, the optimal strategy tends to maintain a larger positive capacity redundancy to cover sudden uncertainties; conversely, it tends to operate close to the demand boundary to save costs. Based on the above problem definition, subsequent chapters will introduce: an anomaly fragment screening mechanism based on prediction residuals, a robust discrimination method based on semi-supervised self-training, and a capacity addition and collaborative release strategy that can be executed online under incomplete information conditions.

[0212] 4.2 Initial Screening of Outlier Fragments Based on Prediction Residues

[0213] To promptly identify sudden load spikes in cloud-native clusters, traditional offline analysis methods face response lag issues in continuous monitoring scenarios. Although the PACE model outputs predictive residuals... While the normal periodic and trend terms in the load have been removed, its statistical characteristics are still affected by two factors: First, different business load bases have different values, resulting in significant differences in residual fluctuation amplitudes, making it difficult to use a uniform absolute threshold for detection; second, monitoring data in the production environment is often accompanied by single-point noise caused by acquisition jitter or instantaneous packet loss. If capacity expansion is triggered directly based on single-point residuals, it can easily lead to frequent oscillations in control actions. Therefore, BRACE introduces a lightweight streaming anomaly detection method in the first stage, using online incremental statistics to replace the high-overhead sliding window batch processing, performing real-time analysis of continuous residual streams with a time complexity of O(1), and quickly identifying suspected sudden start segments.

[0214] To achieve a unified metric across business volumes, it is necessary to analyze the original residuals. Dynamic Z-Score normalization is performed. Considering that traditional sliding window statistics require maintaining a buffer queue and recalculating each time new data arrives, the computational cost increases linearly with the window length, making it difficult to meet the real-time requirements of high-frequency streaming data. This invention introduces a variant of Welford's Online Algorithm, maintaining the statistical benchmark of the residual sequence through recursive updates. Definition and They are time points The residual mean and variance estimators are introduced, along with a forgetting factor. To accommodate the slow drift of the statistical distribution. At time... Obtain new residuals When updating statistics, the following recursive formula is followed:

[0215] (28)

[0216] in, To prevent the use of tiny constants with a denominator of zero, the above recursive formula relies only on the state of the previous time step, without needing to backtrack to historical data. This ensures that even when handling massive concurrent metrics, the system's computational latency remains negligible. Based on real-time updated statistics, time step... Dynamic standardized residuals Defined as Through this standardized transformation, It characterizes the degree of deviation of the current prediction error from the recent historical fluctuation level, thereby eliminating the difference in business volume and enabling the system to handle heterogeneous loads with a uniform dimensionless threshold.

[0217] In obtaining standardized residuals Subsequently, to balance rapid response and noise robustness, this invention designs continuous out-of-bounds triggering logic, aiming to distinguish between transient noise and the initial stage of a real burst. Real flow bursts are physically manifested as a rapid pressure buildup process, rather than an isolated single-point jump. Therefore, the system sets a basic trigger threshold. With minimum number of continuous observations The system maintains a sliding counter. Its state transition logic is as follows:

[0218] (29)

[0219] once achieve The system immediately determines that it is in a potential burst state. At this time, the system does not wait for the burst process to end, but immediately captures a short window segment back from the current moment as a feature snapshot of the burst's rising edge. It then transmits the data in real time to the subsequent semi-supervised classification module for diagnosis. This detection-triggered mechanism strictly controls the physical delay of sudden identification. Within each sampling period, it effectively solves the problem of capacity expansion lag caused by waiting for a complete window in traditional methods.

[0220] To more clearly demonstrate the online streaming processing logic of this process, Figure 6 The diagram depicts the complete data flow from residual input to state update and threshold determination. As shown in the figure, the core of this mechanism lies in the closed-loop state memory module: updating the mean recursively. With variance The system avoids repeated access to and storage of historical raw data, thereby ensuring... Its time complexity enables it to process high-throughput monitoring data streams with extremely low latency.

[0221] To further clarify the engineering implementation details of the above logic, this invention formalizes the streaming screening process into an algorithm as shown in Table 1. This algorithm details how to utilize the forgetting factor. Dynamically update statistics and combine them with continuous counters. Lock down the entry point for emergencies.

[0222]

[0223] From a signal processing perspective, the above screening process essentially constructs a bandpass filter for time-series signals: on the one hand, through dynamic Z-score normalization, the slowly changing drift components and normal random fluctuations captured by the PACE model are filtered out. Effective filtering ensures that the detector is sensitive only to significant structural changes; on the other hand, through The continuity constraint ensures that short-duration high-frequency noise spikes are blocked because they cannot meet the counter conditions, thus avoiding resource jitter caused by false triggers. The candidate fragment set obtained after initial screening... Only abnormal rising edges with a certain strength and persistence are retained, which greatly reduces the frequency of calling subsequent complex models and achieves an initial balance between computational efficiency and detection accuracy.

[0224] 5. Burst detection based on semi-supervised self-training

[0225] 5.1 Temporal morphological feature extraction and pseudo-label construction

[0226] Following the initial residual screening in the previous section, the system extracted a set of candidate anomaly segments from the continuous load stream. However, simply relying on the amplitude exceeding the limits of the standardized residuals is insufficient to accurately distinguish between genuine business bursts and noise interference caused by non-business reasons. To characterize the temporal behavior of candidate segments at a finer granularity, this invention goes beyond simple statistical analysis and constructs a high-dimensional feature vector that includes intensity, morphology, and energy distribution. The aim is to separate different types of anomaly segments in the feature space through multi-view feature mapping. Regarding the first... The duration is candidate fragments The key features extracted in this invention cover three dimensions: steady-state strength, rise rate, and energy structure.

[0227] First, in terms of intensity, considering that single-point peak values ​​are highly susceptible to deviations caused by observation noise, this invention introduces the 95th quantile. As a robust measure of steady-state stress in a segment, it characterizes the level of deviation of the segment over the vast majority of the time; it also incorporates the mean within the segment. These together constitute a subset of intensity features. Secondly, in terms of morphology, genuine burst flows typically manifest as a rapid accumulation of pressure over a short period, while random noise often lacks a clear directionality. Therefore, this invention defines a normalized rising slope. To characterize the load ramp-up pattern, the calculation formula is as follows:

[0228] (30)

[0229] in and These correspond to the residual values ​​at the peak and the initial time, respectively. The slope feature effectively captures the typical dynamic characteristics of a sudden event's steep rise, distinguishing it from slow drift or disordered fluctuations. Finally, in the energy dimension, to identify the sustained impact capability of a sudden event, energy concentration is defined. For segments exceeding the high-bit threshold The proportion of sample energy to total energy, i.e.

[0230] (31)

[0231] This feature can distinguish between high-energy, persistent true bursts and low-energy, sparse transient spikes from the perspective of signal energy, ensuring that only anomalies with sufficient destructive power are marked as bursts.

[0232] After feature extraction, to address the cold start problem in cloud-native scenarios characterized by sudden sample scarcity and lack of real-time manual annotation, this invention employs a heuristic rule-based two-end anchoring strategy to construct an initial pseudo-label set. This strategy leverages prior business knowledge, using samples with extremely high confidence in the feature space as initial anchor points: for samples that simultaneously satisfy… and The fragments are considered as positive sample anchors with extremely strong burst characteristics. For extremely short durations or high energy concentrations Segments below the noise lower bound are considered negative sample anchors. The remaining ambiguous samples that are not explicitly covered by the rules are classified into the unlabeled set. Although this partitioning method is simple, it can provide high-purity initial seed data for subsequent semi-supervised learning, thereby initiating the model's self-training iterative process without human intervention.

[0233] However, directly relying on this set of rules for online detection has significant limitations: rule-based methods essentially define a rectangular hard decision boundary in a high-dimensional feature space. In real production environments, burst traffic patterns are highly diverse and complex, with many atypical bursts existing in the 'critical region,' such as those with a slightly lower ramp rate. However, there is a persistent pressure buildup of highly concentrated energy. While these samples pose a threat to system stability, they are directly filtered out by rigid rule thresholds, resulting in insufficient recall. Therefore, rules can only serve as a means to obtain high-purity seed samples during the cold start phase, and cannot be used as the final judgment criterion.

[0234] 5.2 Confidence-Aware Self-Training Iterative Framework

[0235] After constructing an initial set of high-confidence anchor points, although the purity of the seed samples was ensured by utilizing business priors, this data only covered the typical extreme forms with distinct edges in the feature space. In real cloud-native environments, a large amount of bursty traffic exhibits atypical characteristics. If relying solely on rules, these ambiguous samples located near the decision boundary will be missed. Therefore, the core innovation proposed in this section lies in how to use a small number of textbook-style samples generated by rules as seeds to drive the model to autonomously explore and reconstruct complex nonlinear decision boundaries.

[0236] To address this, the present invention establishes a self-training iterative framework based on confidence perception. This framework does not rely on expensive manual annotation, but instead adopts the Teacher-Student paradigm in semi-supervised learning. Through a closed-loop iteration of pseudo-label generation, sample backfilling, and model retraining, the classifier gradually learns the high-dimensional correlations between features, thereby generalizing the recognition capability from defined anchor regions to unknown and ambiguous regions.

[0237] Let the initial labeled dataset generated by the rules be . (i.e., high-purity anchor points), the initial massive unlabeled dataset is The self-training process is defined as follows: The iterative loop. In the... During each iteration, the algorithm performs the following key steps:

[0238] 1) Base classifier training and probability inference: First, utilize the currently expanded training set. Training base classifier To capture the nonlinear interaction between the intensity, morphology, and energy distribution of burst flows, this invention selects... As a base classifier. For unlabeled pooling. Any unlabeled sample in The classifier outputs a posterior probability estimate of whether it belongs to the burst category:

[0239] (32)

[0240] in For the first The model parameters are learned in each iteration. At this point, the model is no longer just a repeater of rules, but begins to make probabilistic judgments on unknown samples based on the learned feature patterns.

[0241] 2) Confidence Screening Based on Dual Thresholds: In the early stages of iteration, the model performance is unstable. Directly using all predicted results as true values ​​for backfilling can easily introduce noise and lead to error accumulation. Therefore, the algorithm introduces an asymmetric dual threshold screening mechanism to generate pseudo-labels. The selection function is defined. False labels are only assigned when the predicted probability exhibits extremely high certainty. :

[0242] (33)

[0243] in and These are the high and low position confidence thresholds for distinguishing between sudden and non-sudden events. This mechanism explicitly filters intermediate samples in the decision-making state, ensuring that the pseudo-labels in each round of expansion have extremely high confidence.

[0244] 3) Sample Transfer and Boundary Expansion: Based on the selection function, extract a high-confidence sample subset from the unlabeled pool for this round. :

[0245] (34)

[0246] Subsequently, a dataset update operation is performed, incorporating the newly labeled samples into the training set and removing them from the unlabeled pool:

[0247] (35)

[0248] Through the aforementioned closed-loop feedback mechanism, with each iteration... With the increase of training set The sample distribution gradually spreads from the central region to the peripheral region of the feature space. Essentially, this process involves the model, under the constraint of dual thresholds, utilizing the learned feature patterns to understand and accept samples that initially do not meet the rule definitions but whose inherent patterns are highly consistent with the sudden changes. Ultimately, the model will converge to a robust discriminator that possesses both high accuracy of the rules and high generalization ability of deep models.

[0249] To visually illustrate the above iterative process, Figure 7 The document demonstrates a confidence-aware self-training framework. This framework forms a closed-loop feedback system: Stage 1 initializes the teacher model using anchor data; Stage 2 uses a dual-threshold filter to transform the student model's predictions of unlabeled data into high-confidence pseudo-labels; Stage 3 refills the model with new knowledge and triggers a new round of training until the model converges or reaches the maximum number of iterations.

[0250] The specific iterative training logic is shown in Table 2. The algorithm explicitly defines the training process from the unlabeled pool. To training set The sample migration process, and the pseudo-label-based incremental processing. The convergence and stopping mechanism ensures the stability and controllability of the offline training process.

[0251] It is worth emphasizing that the classifier learned through this offline iterative framework no longer bases its decisions on simple numerical values, but rather on the morphological features of the residual sequence. Since sudden traffic surges at different times are essentially short-term, high-energy impacts on system resources, this morphological feature-based discrimination logic possesses inherent drift-resistance. Therefore, once offline training is complete, the discriminator's parameters can be frozen and deployed to the online inference engine. Its extremely low latency processing of real-time data streams ensures long-term recognition stability while completely avoiding the risks of model degradation and catastrophic forgetting that may be introduced by online learning.

[0252] 5.3 Methods for Determining Emergency Situations

[0253] Following the streaming initial screening mechanism in Section 5.2, when the continuous out-of-bounds counter reaches the trigger condition, the system does not immediately classify it as a true burst. Instead, it first sends the candidate anomaly window corresponding to the current moment into the semi-supervised discriminant model for further confirmation. The purpose of this is to combine the rapid initial screening in the previous stage with the fine identification in the subsequent stage, thereby reducing the risk of misjudgment caused by relying solely on residual out-of-bounds errors while ensuring the timeliness of detection.

[0254] In the online discrimination phase, the candidate window undergoes feature extraction to construct input features consistent with those used in the offline training phase, considering aspects such as intensity, morphology, and energy structure. These features are then fed into a robust classifier whose parameters have been trained and frozen. The classifier outputs the confidence score that the window belongs to the burst category, thereby determining the state of the current candidate event. Since the model has learned the distribution differences between typical burst and non-burst segments in the high-dimensional feature space through offline self-training, its discrimination criteria are no longer limited to whether a single image value exceeds the bounds. Instead, it comprehensively considers the overall temporal morphology and evolutionary characteristics of the abnormal segments, thus enabling more effective differentiation between genuine business bursts, random noise interference, and slow load drift.

[0255] To avoid frequent fluctuations in classification results near the decision boundary, the online discrimination rule employs a dual-threshold confidence mechanism: when the classifier output reaches the high confidence interval, the system determines the current candidate window as a sudden event; when the output falls into the low confidence interval, it is determined to be a non-sudden event; when the result is in the intermediate uncertainty region, a sudden alarm is not immediately triggered, but the current state is maintained or subsequent samples are observed to suppress false triggers and state fluctuations caused by boundary samples. Through this two-stage discrimination method of initial screening triggering - model confirmation - threshold stabilization, the system can reliably identify sudden events with a relatively small observation delay.

[0256] Once the model provides its first confirmed diagnosis of an emergency, the system designates that moment as the official trigger point for the emergency and transmits the result to the capacity decision module to drive subsequent transient capacity additions and coordinated release control. Thus, this invention completes the entire identification chain from anomaly fragment detection and candidate sample screening to final confirmation of the emergency, providing reliable event input for subsequent resource response.

[0257] 5.4 Transient Capacity Attachment and Coordinated Release Strategies

[0258] In the preceding chapters, the system has achieved rapid identification of sudden loads through streaming residual screening and semi-supervised robust discrimination. However, for BRACE, simply determining whether a sudden load has occurred is insufficient to complete the resource management loop; a more critical issue is how quickly and to what extent the system should add computing resources after a sudden load is confirmed, in order to control SLO default risk while avoiding oversupply. Unlike the baseline capacity planning under normal loads faced by PACE in Chapter 4, this section focuses on the emergency additional capacity decision-making problem under sudden load shocks, that is, further determining the injection and withdrawal methods of short-term compensation capacity on top of the baseline replica count already given by PACE.

[0259] Based on the joint optimization objective of default risk and resource cost established in Section 5.1, resource disposal in the initial stage of a sudden event exhibits significant online decision-making characteristics: on the one hand, the system must act quickly when information is insufficient; on the other hand, the intensity, duration, and subsequent evolution path of the sudden event are all uncertain, making it difficult to directly solve for the globally optimal additional capacity sequence in engineering. Therefore, this invention does not attempt to solve the complete joint optimization problem in real time, but instead constructs an online approximation strategy for sudden events: when a sudden event is triggered, the capacity gap is conservatively estimated based on the maximum deviation within the residual window and mapped to a one-time additional capacity above the baseline; after the sudden event ends, the emergency capacity is gradually withdrawn through a linear step-by-step release mechanism, thereby achieving smooth coordination between the BRACE link and the PACE link.

[0260] 1) Aggressive capacity mapping based on maximum residual

[0261] At any moment Once the flow cytometer confirms a burst event, the system first needs to estimate the additional pressure the burst places on resource supply. This is based on the predicted values ​​output by the PACE model. Having already undertaken the normal capacity planning functions corresponding to the periodic and trend terms, BRACE does not recalculate the total capacity demand at this stage, but only compensates for the instantaneous gap not covered by the normal model. Since emergencies often exhibit steep increases and strong non-stationary fluctuations in their initial stages, using average residuals or smoothing statistics can easily underestimate resource demand at the most dangerous moment. To improve the conservatism of emergency decision-making, this invention selects an emergency detection window. The largest unstandardized residual observed within the internal observations is used as an estimator of the load gap:

[0262] (36)

[0263] in This represents the triggering moment of the current emergency. This estimate does not attempt to characterize the average intensity of the entire emergency, but rather prioritizes covering the most destructive instantaneous peak pressure in the initial stages, thus providing a safety boundary for subsequent additional capacity calculations.

[0264] After obtaining the load gap estimate, the system needs to map it from the traffic space to the resource space. This is based on the default risk penalty in Section 5.1. With resource costs To address the trade-offs, this invention introduces an aggressive expansion factor. This coefficient reflects the business's sensitivity to SLO (Service Level Agreement) default: a larger value indicates that the system is willing to pay higher short-term resource overhead for lower default risk when the business has higher requirements for service stability. This is combined with the single container's processing capacity limit. ,time Additional capacity The calculation is as follows:

[0265] (37)

[0266] The above mapping process reflects BRACE's emergency response capabilities. Firstly, by dividing the residual gap by the single-instance capacity, the system converts the additional load pressure into the number of additional instances required, and then uses rounding up. First, ensure that resource allocation meets the discrete replica granularity constraint; second, introduce a maximum additional capacity quota. The system has set an upper bound for burst response protection to prevent unconstrained expansion of additional capacity due to abnormal monitoring or misjudgment. Ultimately, the total target number of replicas issued by the execution layer... The baseline replica count provided by PACE is combined with the additional capacity output by BRACE, i.e.

[0267] (38)

[0268] This superposition relationship indicates that BRACE does not replace the baseline planning function of PACE, but rather serves as an independent abnormal compensation link to provide short-term enhancement to baseline capacity during emergencies. Even if the baseline prediction of PACE has not yet had time to rise, BRACE can still rely on additional capacity to quickly absorb abnormal pressure, thereby preventing the system from being exposed to significant capacity gaps in the early stages of an emergency.

[0269] 2) Linear attenuation and dual-link cooperative mechanism

[0270] After the surge is identified and additional capacity is added, another key issue is: how should the added capacity be deactivated when the surge weakens or even disappears, in order to avoid both resource waste and the system falling back into a state of undercapacity? If the additional capacity is directly deactivated after the detection signal disappears... If the load drops to zero, the total system capacity will instantly revert to the old baseline level. If the current surge is not a short-duration pulse, but a precursor to a shift in the load structure, this drastic pullback will cause the system to experience severe under-configuration on the new load platform, thus triggering a secondary overload.

[0271] To address this, this invention designs a linear step-by-step release mechanism. Its core idea is not simply slow capacity reduction, but rather to retain a brief emergency buffer window after the burst ends, allowing the PACE normal link time to re-sensitize and adapt to the new load state. Specifically, when the system exits the burst state, the additional capacity no longer fluctuates instantly with the residual signal, but instead enters a decay phase. Definition For a single-step release rate, the additional capacity decreases according to the following rule:

[0272] (39)

[0273] The purpose of this mechanism is to establish a smooth handover interval between the two links. During this period, the additional capacity of BRACE output gradually decreases, while the baseline prediction and baseline replica count of PACE continue to be updated online, thereby avoiding abrupt switching of resource control between the two modules.

[0274] If the current event is an occasional, short-lived burst, then during the additional capacity decay process, the predicted value of PACE will be... Compared with baseline copy number If the system essentially maintains its original level, the total system capacity will naturally return to its normal operating point as the load decreases, thus avoiding the prolonged maintenance of unnecessary redundant resources after the burst ends. Conversely, if the burst actually corresponds to a continuous increase in the load baseline, meaning the burst gradually evolves into a new normal level, then... As the degradation slows, the drift modeling mechanism in the PACE link will gradually absorb this change and improve baseline predictions. Compared with baseline copy number The capacity increases synchronously. At this point, the additional capacity released by BRACE will be smoothly replaced by the newly formed baseline capacity of PACE, and the total system capacity will not experience a significant drop.

[0275] Therefore, the exit process of BRACE is not essentially a separate reduction in capacity, but rather a process of gradually reducing additional capacity to create a buffer time for the baseline update of PACE. During this phase, the two links form a coordinated emergency backup and baseline takeover relationship, enabling the system to smoothly transition between emergency response states and normal steady-state control, avoiding common problems in traditional strategies such as expansion followed by reduction, reduction followed by expansion, and capacity gaps.

[0276] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0277] To verify the detection performance and robustness of the BRACE strategy burst discrimination model in this invention under scarce labels, an example was conducted. This example focuses on evaluating whether the model can effectively converge using a semi-supervised self-training framework with only a small number of high-confidence anchor points as initial seeds, and whether it can achieve high recall for burst traffic while ensuring low false positives.

[0278] Burst traffic identification is essentially a binary classification problem, which determines whether the current load residual sequence belongs to the burst category. To comprehensively evaluate the discriminative performance of the BRACE model under scarce labels, the experiment counted the number of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) based on the confusion matrix, and selected accuracy, precision, recall, and F1 score as the core evaluation metrics.

[0279] Accuracy is used to intuitively measure the overall correctness of the model's judgments across all test samples, reflecting the model's global performance. Precision describes the proportion of genuine burst events among all samples judged as bursts by the model; this metric is directly related to resource costs, and higher precision means less ineffective scaling and wasted computing resources due to misjudgments. In cloud-native elastic scaling business scenarios, recall is considered the most critical core metric. It measures the proportion of genuine burst events successfully identified by the model out of all actual burst events. Since false negatives mean that the system fails to detect and trigger scaling in time when the load surges, this will directly lead to request backlogs, latency spikes, and even service quality (SLO) breaches, the consequences of which are far more serious than the small amount of idle resources caused by false positives. Therefore, the experiment focuses on examining how the model can maximize precision while pursuing 100% recall. Finally, the F1 score, as the harmonic mean of precision and recall, is used to seek a balance between the two and comprehensively evaluate the robustness of the model.

[0280] From the perspective of the self-training iteration process, the model performance shows a significant increase with the increase of training rounds and eventually tends to stabilize. For example... Figure 8 As shown, in the initial stage, due to the limited number of available labeled samples, the model's discriminative ability is weak, and the Precision, Recall, and F1 scores are generally at a low level. As high-confidence pseudo-labeled samples are continuously added to the training set, all three metrics continuously improve. Precision increases the fastest and stabilizes at a high level in the later stages, indicating that the model gradually develops a strong ability to reject non-sudden samples during self-training, effectively suppressing the spread of false positives. Meanwhile, Recall also gradually increases with iterations, indicating that the model's coverage of sudden samples is continuously enhanced. Although the improvement in recall is relatively more gradual, it ultimately remains at a high level, indicating that semi-supervised self-training not only improves the model's fitting ability to known samples but also enhances its ability to identify unknown sudden patterns. The F1 score tends to stabilize in the later stages of iteration, indicating that the model has reached a relatively stable trade-off between precision and recall, and the training process as a whole exhibits good convergence.

[0281] Figure 9 The confusion matrix on the test set is presented. As shown in the figure, the model's identification results for non-burst samples are TN=701 and FP=19, while for burst samples, the results are TP=375 and FN=105. Based on the confusion matrix, the model's accuracy on the test set is 89.7%, precision is 95.2%, recall is 78.1%, and F1 score is 85.8%. These results demonstrate that the most prominent feature of the current model is its strong false alarm control capability: out of 720 real non-burst samples, only 19 were misclassified as bursts, corresponding to a specificity of approximately 97.4% and a false alarm rate of approximately 2.6%. This indicates that the model can distinguish real bursts from false anomalies caused by acquisition jitter, short-term spikes, or local fluctuations in most cases, thus avoiding unnecessary expansion overhead due to false triggers. On the other hand, the model's recall rate for burst samples reaches 78.1%, indicating that it can identify most real burst events, but a certain proportion of samples are still missed. This also shows that under the condition of scarce labels and complex boundaries, the model tends to adopt a relatively conservative discrimination strategy, that is, to prioritize high accuracy and then improve the coverage of sudden samples as much as possible.

[0282] To further explain the source of the above classification results, this invention performs PCA dimensionality reduction visualization on the discriminant features, and the results are as follows: Figure 10As shown in the figure, non-burst samples are mainly concentrated in the left region, while burst samples are mainly distributed in the right region, indicating that the constructed features have good discriminative ability in the main direction. Specifically, the first principal component (PC1) explains 78.0% of the variance, indicating that the main separable information of the two classes of samples is concentrated in this direction; the second principal component (PC2) explains 8.8% of the variance, reflecting more the intra-class dispersion and local perturbation differences. The background color in the figure indicates the approximate discriminative region of the model in the two-dimensional projection space. It can be seen that the decision boundary formed by the model generally expands along the main distribution direction of the two classes of samples, achieving relatively clear class division in most areas.

[0283] However, in the transition region where PC1 approaches 0, there is still some overlap between the two types of samples. These samples near the boundary often exhibit characteristics such as low burst intensity, short duration, or significant local fluctuations without forming a sustained overall impact, making them more likely to become the main source of misclassification and missed classification. The unlabeled samples in the figure form a relatively continuous distribution band between the two types of labeled samples, indicating that the self-training process does indeed utilize unlabeled data to fill the gaps between the original seed samples, allowing the model to gradually correct the discrimination boundary under limited label conditions and improve overall generalization ability.

[0284] In summary, the semi-supervised self-trained burst discrimination model in BRACE demonstrates good training stability and testing performance under scarce label conditions. Its advantages are mainly reflected in two aspects: first, through continuous iteration, it achieves performance improvement from low-quality initial discrimination to stable convergence; second, in the final testing phase, it achieves high accuracy and good comprehensive recognition capabilities, enabling it to identify most real burst events while controlling false positives.

[0285] To comprehensively evaluate the resource supply optimization capability of the BRACE strategy under sudden traffic surges, this invention selects a load in the dataset that includes periodic characteristic traffic and high-frequency pulse bursts. Regarding system physical constraints, the upper limit of the processing capacity μ of a single container instance under the SLO (Solution Time Limit) is set to 10 req / s, and the SLO response time threshold L_SLO is set to 0.2 seconds. Considering the cold start problem commonly encountered in cloud-native environments, the experiment models the effective latency from decision issuance to instance readiness as a random variable with a mean of 25 seconds and a jitter range of ±3 seconds to reproduce the real scheduling lag effect.

[0286] In terms of optimization objectives and comparison benchmarks, the experiment introduced a dual consideration of resource cost and default risk. The resource billing model adopts the typical pay-as-you-go model of public cloud, and sets the operating cost β of a unit container instance to 0.0014 yuan / minute. In order to verify the superiority of the method of the present invention, the experiment selected three representative strategies as benchmarks: Kubernetes native HPA strategy, AdaFrame

[24] strategy and BAScaler

[28] strategy. The target utilization threshold of all baseline strategies was uniformly set to 60%. In contrast, the BRACE strategy proposed in this invention uses streaming Z-Score statistics for burst perception, sets the trigger threshold to 2.0σ, and sets the single-step decay factor of additional capacity to 0.55, aiming to achieve millisecond-level response in the early stage of the burst and rapid cost recovery after the burst ends.

[0287] To verify the actual performance of the BRACE strategy under the joint optimization objective, this invention combines simulation experiments to generate the following result graph for in-depth analysis. Experimental results show that BRACE successfully overcomes the inherent contradiction of high performance inevitably leading to high cost in traditional elastic scaling when dealing with sudden traffic surges, significantly reducing SLO default rates while effectively controlling the cumulative resource overhead.

[0288] 1) Analysis of Emergency Response and SLO Assurance Capabilities

[0289] In the burst experiment in this section, in order to more intuitively demonstrate the arrival of burst requests, capacity follow-up lag, and supply-demand deviation within the startup window, RPS (Requests Per Second) is used as a horizontal comparison indicator.

[0290] Figure 11 The presentation visually illustrates the capacity tracking performance of four strategies under three-pulse bursts. It's evident that Kubernetes HPA, relying on an exponentially weighted moving average to smooth metrics, exhibits a significant lag effect, remaining under-capacitated for the first 25-30 seconds after the burst arrives, causing the capacity supply curve to shift significantly to the right compared to the load demand curve. While AdaFrame introduces a derivative term for prediction, its prediction gain is often insufficient to cover the 25-second cold start latency window when faced with steep, non-linear rising edges.

[0291] In contrast, BRACE demonstrated exceptional agility. Thanks to the millisecond-level triggering mechanism of streaming Z-Score, BRACE identified residual anomalies the instant of a burst start and immediately calculated an aggressive additional capacity Δn_k. This detection-and-execute mechanism effectively offset the time window loss caused by cold starts, allowing BRACE's capacity curve to climb almost synchronously with the burst rising edge, rapidly filling the supply-demand gap.

[0292] Figure 12 The performance benefits of this response speed were further quantified. At the three peak bursts, HPA and AdaFrame both experienced instantaneous default traffic of 500-800 rps, meaning a large number of requests timed out while queuing. In contrast, BRACE's default traffic was suppressed to an extremely low level, with only a small overflow in the first few seconds of the burst, quickly returning to zero. This demonstrates that BRACE's aggressive scaling strategy successfully prevented a service quality avalanche and effectively guaranteed user experience in high-concurrency scenarios.

[0293] 2) Cost performance and analysis under emergency scenarios

[0294] This section focuses on examining BRACE's effectiveness in controlling resource overhead during burst scenarios. Since the billing price per container instance remains constant, differences in capacity adjustment strategies ultimately translate into cost differences. Therefore, this invention first compares the performance of various methods in terms of instantaneous cost rate and cumulative expenditure to assess whether BRACE can effectively suppress additional resource overhead while ensuring burst response capabilities. Subsequently, it explains the reasons for its cost advantage by considering the changes in the number of active instances, resource utilization, and idle capacity.

[0295] First, looking at the changes in instantaneous cost rates, BRACE exhibited a relatively stable cost evolution process across multiple sudden ranges. For example... Figure 13 As shown, the cost rates of HPA, AdaFrame, BAScaler, and BRACE all increase with load, but there are significant differences in peak magnitude, duration, and rate of decline among the methods. HPA maintains a high cost rate even after several bursts, exhibiting a noticeable high-level tail, indicating that it retains a significant amount of additional resources after the load decreases. BAScaler shows sharper cost peaks in some burst intervals, reflecting its more aggressive expansion. AdaFrame's cost changes are generally smooth, but it still experiences high instantaneous expenditures near major peaks. In contrast, while BRACE's cost rate also rises rapidly during bursts, the peak is more restrained overall, and the rate of decline after the burst ends is faster, without significant prolonged high-cost stagnation. This suggests that BRACE does not simply reduce costs through conservative supply, but rather aims to minimize the duration of high-cost states while ensuring burst compensation capabilities, thereby achieving more balanced cost control.

[0296] To further compare overall costs, Figure 14The cumulative cost difference curves of each baseline method relative to BRACE are presented. Here, the cumulative cost difference is defined as the cumulative cost of the baseline method minus the cumulative cost of BRACE. Therefore, when the curve is above zero, it indicates that BRACE has achieved a cumulative cost advantage relative to the corresponding method. The results show that the three curves generally trend upwards over time, indicating that BRACE gradually establishes a more significant cumulative cost advantage in continuous burst scenarios. Specifically, compared to HPA, BRACE formed a relatively stable positive cumulative cost difference at the end of the experiment, indicating its superior long-term cost control effect; compared to BAScaler, BRACE's cumulative advantage is more significant, indicating that while aggressive scaling can improve short-term response capabilities, it also more easily leads to additional cost accumulation; compared to AdaFrame, BRACE still maintains a positive advantage, but the difference is relatively smaller, indicating that AdaFrame itself already has a certain cost control capability, while BRACE further reduces cumulative cost on this basis. It is worth noting that some curves briefly approached zero or even fluctuated slightly during certain periods. This indicates that BRACE's advantage does not lie in its absolute optimality at all instants, but rather in its continuous cumulative effect over multiple bursts. In other words, BRACE's cost-benefit profile is more of an overall advantage in a dynamic scenario than a random result at a particular local moment.

[0297] To explain the reasons for this cumulative cost advantage, Figure 15 The changes in the number of active containers for each method are further presented. The results show that BRACE can promptly increase the number of replicas to provide necessary compensation during a burst, but its capacity peak is generally lower than the extreme highs of HPA, AdaFrame, and BAScaler, and its rollback process after the burst is smoother. HPA exhibits a long high-level tail after multiple burst intervals, indicating that its scaling down is relatively lagging; BAScaler shows more obvious instantaneous spikes and jitters in several intervals, demonstrating strong aggressive behavior; AdaFrame's overall change is relatively smooth, but it still maintains a high capacity near the main peaks. In contrast, BRACE's adjustment characteristics are more in line with the strategy logic of on-demand compensation and timely exit, that is, providing necessary additional capacity during the burst phase and quickly rolling back ineffective redundancy after the event subsides. Since the cost rate is essentially determined by the number of active container instances, this more restrained and smoother capacity evolution is a key reason why BRACE can reduce instantaneous cost peaks and form a cumulative cost advantage.

[0298] In addition to capacity trajectories, changes in resource utilization further explain the differences in cost performance among various methods. For example... Figure 16As shown, HPA's utilization drops significantly to a lower level during the recovery phase after some bursts, indicating that it retains a large amount of resources even after the load has subsided, resulting in additional expenses. BAScaler and AdaFrame, on the other hand, rapidly rise to near-saturation levels during certain periods. While this can increase short-term load capacity, it also means that their capacity adjustment is more aggressive, easily leading to subsequent oversupply or pullback fluctuations. In contrast, BRACE can quickly return to a more reasonable utilization range in most periods, without remaining in a significantly low or persistently high-pressure extreme state for extended periods. This indicates that BRACE's additional capacity is not an unconstrained accumulation, but rather a better match to the burst demand itself, thus avoiding the inefficient operation of resources being over-invested even after the load has subsided. In other words, BRACE's cost advantage is not simply achieved by reducing the number of Pods, but is based on a capacity supply that better matches the actual load changes.

[0299] Furthermore, Figure 17 The given results on idle capacity changes provide more direct evidence for this point. Idle capacity reflects the positive difference between the system's available capacity and the actual load, thus providing a more intuitive indication of how much unused redundant capacity remains after a burst. As shown in the graph, HPA generated high idle capacity peaks after several burst intervals, with a slow decline, indicating significant resource retention after the high load subsided. AdaFrame also exhibited significant redundancy accumulation in certain phases after the main peak. BAScaler, due to its more aggressive expansion, formed sharper idle capacity peaks in several intervals. In contrast, BRACE's overall idle capacity is lower and falls back more quickly after most burst events, indicating that its additional resources exit more promptly after emergency compensation. This lower level of redundancy retention directly reduces the duration of ineffective resource occupation, thus constituting a significant source of BRACE's cumulative cost advantage.

[0300] In summary, the results show that BRACE demonstrates superior cost control capabilities in emergency scenarios. Firstly, at the instantaneous level, BRACE avoids prolonged periods of high costs, resulting in relatively stable cost ratios. Secondly, at the overall level, BRACE exhibits varying degrees of cumulative cost advantages compared to HPA, AdaFrame, and BAScaler. Thirdly, from a mechanism perspective, this advantage primarily stems from more controlled capacity peaks, more timely additional capacity rollback, more rational resource utilization, and lower idle capacity retention. Therefore, BRACE not only provides necessary resource compensation during emergencies but also effectively suppresses ineffective redundancy after the event subsides, thus achieving superior cost-benefit performance in continuous dynamic scenarios.

[0301] In cloud-native container cluster resource management scenarios within a computing power network context, business loads typically exhibit a complex combination of cyclical patterns, slow-moving drift, and high-frequency disturbances / bursts. If a passive scaling strategy relying solely on threshold triggers, or if the prediction module and control strategy are separated, is prone to false triggers and frequent jitter under noise disturbances, increasing control plane pressure, amplifying instance start-up and shutdown overhead, and introducing cold start lag risks during load ramp-up. To address the problem of rapidly expanding capacity gaps and escalating SLO default risks under sudden traffic surges, this invention proposes the BRACE (Burst Recognition and Adaptive Capacity Expansion) burst identification and adaptive expansion method: at the burst detection layer, the prediction residual of PACE is used as the core signal, combined with adaptive thresholds, morphological constraints, and pseudo-label semi-supervised self-training to improve cross-scenario robustness; at the capacity handling layer, the detection results are mapped to rapid additional expansion above the baseline, and a step-by-step removal mechanism suppresses oscillations and cost waste during the fallback phase.

[0302] This invention achieves coordinated optimization of stability, responsiveness, and cost under complex fluctuating loads by combining the normal predictive control link of PACE with the abnormal emergency handling link of BRACE. Experiments have verified the effectiveness of the proposed method in reducing ineffective scaling and improving the reliability of emergency identification.

[0303] In the description of this invention, unless otherwise stated, "multiple" means two or more; the terms "upper," "lower," "left," "right," "inner end," "rear end," "head," and "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0304] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0305] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying burst loads and scaling additional capacity in a computing network, characterized in that, Includes the following steps: Step 1: A dual-branch collaborative resource management framework based on load decomposition; By perceiving and analyzing real business loads from multiple perspectives, it was found that most online loads are more suitable to be represented as a composite structure of periodic changes + slow change trends + high-frequency fluctuation residuals. Step 2: Divide the resource management process into a prediction-driven normal capacity planning link and an anomaly-driven emergency capacity expansion link, and connect the two links through prediction residuals. Step 3: Using the prediction residuals of the normal link output as input for abnormal disturbances, the improved Welford online statistical algorithm is used in the detection layer to achieve low-latency dynamic standardization and preliminary screening of abnormal segments. In the discrimination layer, a semi-supervised self-training iterative framework based on pseudo-labels is introduced. In the capacity handling layer, a rapid additional expansion and step-by-step release mechanism is combined to achieve rapid capacity compensation and smooth recovery in sudden scenarios.

2. The method for identifying and scaling additional capacity in a computing network as described in claim 1, characterized in that, The process involves multi-faceted perception and analysis of real business load: A. Statistical distribution characteristics The most basic quantity is the mean, which describes the average load level within the window: （1） Based on this, the maximum and minimum values , and range It is often used to reflect peak pressure and value range, making it easy to identify load objects that are generally high or have prominent peak values; To measure the intensity of load fluctuations around the mean, the standard deviation is typically calculated. : （2） Considering that different objects may have different mean levels, only... Often incomparable, the coefficient of variation (CV) is frequently used in practice to characterize relative fluctuations. Here, we define CV as follows: （3） in It is a numerically stable term; generally speaking, The larger the value, the stronger the fluctuation at the unit load level, the worse the load stability, and the easier it is to induce frequent capacity adjustments at the control level. When a load series exhibits peaks or tails, mean-variance statistics are easily affected by extreme points. Therefore, quantiles are often introduced to more robustly describe the distribution location: Let quantiles be... And the robust dispersion is characterized by the interquartile range: （4） Quantiles and IQR can eliminate the interference of a small number of extreme points to a certain extent. They are often used to compare the typical operating range and dispersion of different load objects, providing a basis for subsequent anomaly detection or threshold setting. B. Trend and Slow-Change Characteristics A typical approach to characterizing trends is to fit a window sequence using a univariate linear model: （5） in This represents the average rate of change of the load within the window over time. The intercept is... The remaining term not explained by the trend; slope This usually corresponds to a sustained increase in load. The corresponding load continues to decline; This quantity is often used in load analysis to distinguish between monotonic evolution and oscillating types without a clear direction, and to provide directional information for subsequent capacity planning; Let the correlation coefficient of the regression fit be Then the trend strength index can be defined as （6） Building upon trend fitting, a crucial next step is detrending, which involves removing the portion of the original sequence that can be explained by the trend, resulting in the residual sequence. （7） The residual standard deviation can then be used to measure the magnitude of the remaining volatility after removing slow variations: （8） This quantity has direct meaning in load-aware analysis: if The load remains relatively large, indicating that fluctuations within the window are primarily driven by rapid ups and downs, spikes, or noise; trend modeling alone is insufficient to explain the load. A significant decrease indicates that the slow-changing trend accounts for a larger proportion of the load explanation, and the load has a stronger structural nature. C. Features related to the rate of change; D. Time-related features; E. Frequency domain periodicity characteristics; F. Complexity-related features.

3. The method for identifying and scaling additional capacity in a computing network as described in claim 2, characterized in that, The change rate type feature: In load-aware analysis, differential classification features are typically introduced to measure the magnitude and direction of local changes in a sequence. The most commonly used is the first-order difference sequence: （9） based on Three intuitive quantities can be obtained: its mean The standard deviation represents the overall direction of change within the characterizing window. The mean absolute difference characterizes the stability of the rate of change. This depicts the typical amplitude of the jump; Define the number of ascents and the number of descents: （10） And can be further used in proportion Describe whether rising or falling is dominant within the window.

4. The method for identifying burst loads and scaling additional capacity in a computing network as described in claim 2, characterized in that, The time-related features Lag Autocorrelation can be defined as: （11）。 5. The method for identifying and scaling additional capacity in a computing network as described in claim 2, characterized in that, The frequency domain periodicity class feature: After removing the mean from the window sequence, its Discrete Fourier Transform (DFT) can be constructed. （12） With amplitude spectrum Describe the intensity of each frequency component; locate the main peak frequency within the positive frequency range: （13） （14）。 6. The method for identifying burst loads and scaling additional capacity in a computing network as described in claim 2, characterized in that, The complexity class features: Sample entropy is a commonly used metric in time series complexity analysis. It reflects the reproducibility of local patterns by comparing the pattern-matching probabilities of sequences under different embedding dimensions; let the embedding dimension be... The construction length is Template vector: （15） The similarity between different templates is measured using some distance metric (commonly Chebyshev distance), and the statistical distance is less than the tolerance. The number of matching pairs yields the dimension. Matching ratio below And then expand to dimensions get ; Sample entropy is defined as: （16）。 7. A system for identifying and adding capacity scaling of burst loads in a computing power network, implementing the method for identifying burst loads and adding capacity scaling in a computing power network as described in any one of claims 1-6, characterized in that, The system includes: The perception and analysis module is used for a dual-branch collaborative resource management framework based on load decomposition. By perceiving and analyzing real business loads from multiple perspectives, it is found that most online loads are more suitable to be represented as a composite structure of periodic changes + slow change trends + high-frequency fluctuation residuals. The partitioning module is used to divide the resource management process into a prediction-driven normal capacity planning link and an anomaly-driven emergency capacity expansion link, and connect the two links through prediction residuals. The filtering module uses the prediction residuals of the normal link output as input for abnormal disturbances. In the detection layer, it uses an improved Welford online statistical algorithm to achieve low-latency dynamic standardization and preliminary screening of abnormal segments. In the discrimination layer, it introduces a semi-supervised self-training iterative framework based on pseudo-labels. In the capacity handling layer, it further combines a rapid additional expansion and step-by-step release mechanism to achieve rapid capacity compensation and smooth recovery in sudden scenarios.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for identifying burst loads and scaling additional capacity in a computing network as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for identifying burst loads and adding capacity scaling in a computing network as described in any one of claims 1-6.

10. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the burst load identification and additional capacity scaling system in the computing power network as described in claim 7.

Images